Picture a two person team with a fine-tuned model that already has paying users.
They ask their cloud provider for more GPU capacity this quarter.
The answer is a wait, with no date attached.
Amazon took delivery of roughly $5 billion of Nvidia GB300 server racks in August, and internal emails seen by Business Insider say about $3 billion more is expected every month.
Some of those racks sit in intermediary warehouses because the AWS sites are not finished.
Some US facilities do not have the power to switch the servers on.
One Amazon team wrote that there was no clear forecast for installing all the GPUs the company was acquiring each month.
Amazon's spokesperson called storing finished racks in a warehouse best business practice.
Trailing free cash flow turned negative in the second quarter, with capex heading toward $220 billion this year.
The real constraint is the building and the grid behind it.
So the capacity that team is waiting on is often sitting idle, bought and stored while the site waits for a power feed.
That changes the decision.
Running a 7B model or a fine-tune does not require a megaproject.
RTX 3090 compute at $0.15/hr and storage at $2.49/TB per month are available now, and Buy Now Pay Later puts zero upfront with payment after one month.
What is the longest stretch you have spent between requesting cloud GPU capacity and actually running your first job?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w43&utm_content=evening
Originally posted on linkedin_personal